Maximizing the Value of Claude Code Sessions

Agent-based coding tools like Claude Code can vary significantly in token consumption and cost depending on how tasks are approached. The key strategies for efficient session management are as follows.

Key Optimization Strategies

• Run /clear: Execute this between tasks to remove unnecessary previous context, thereby reducing token usage.
• Use @-mentions: Instead of typing file names directly, mention files with @ to save resources (Read calls or Search) that Claude would otherwise spend locating the files.
• Run /compact: Summarize the conversation before the prompt cache expires to reduce costs.
• Check /context: Review the content loaded at the start of the session (such as CLAUDE.md, MCP tool definitions, etc.) and remove unnecessary elements.
• Pre-specify settings: Set the model and Effort level before starting the session. Changing them mid-session can break the prompt cache and increase costs. (For simple tasks, you can disable thinking tokens by setting MAX_THINKING_TOKENS=0.)

Cost and Token Mechanisms

• Input vs. Output Tokens: Output tokens are generated one by one by the model, occupying the GPU for longer, making them approximately 5x more expensive than input tokens.
• Prompt Caching: For requests starting with the same tokens, the server maintains state, and the cost of reading from the cache is about 0.1x the standard input cost.
• Model Selection: Optimize costs by using larger models for complex problems and smaller models for simple tasks.